Improved Superpoint with Feature Transformer for Remote Sensing Image Registration

Remote sensing image registration plays a significant role in various fields such as disaster monitoring, response, agriculture and forestry management. Registration of remote sensing images is more complicated than natural images due to the larger coverage area and being affected by factors such as the atmosphere, clouds, and sensor noise. To address the limitations of the SuperPoint method commonly used for natural image registration in adapting to large lighting variations or repetitive patterns, this paper introduces a lightweight network to enhance the keypoint descriptors in SuperPoint. The network takes the original descriptors and geometric properties of keypoints as inputs and employs a descriptor enhancement stage and a spatial context enhancement stage to enhance the descriptors. The results indicate that, compared with the SuperPoint approach, the method we optimized shows better performance when tested on Landsat-7 and WorldView-3 images captured at different time periods.